Moradpour, Shohreh ， Zhao, Shuai ， Entezari, Mojgan ， Ayoubi, Shamsollah ， Mousavi, Seyed Roohollah
2025-10-17 REVUE INTERNATIONALE DE GEOMATIQUE 2025 34(卷), null(期), (809-829页)
Accurate spatial prediction of soil organic carbon (SOC) and soil inorganic carbon (SIC) is vital for land management decisions. This study targets SOC/SIC mapping challenges at the watershed scale in central Iran by addressing environmental heterogeneity through a random forest (RF) model combined with bootstrapping to assess prediction uncertainty. Thirty-eight environmental variables-categorized into climatic, soil physicochemical, topographic, geomorphic, and remote sensing (RS)-based factors-were considered. Variable importance analysis (via) and partial dependence plots (PDP) identified land use, RS indices, and topography as key predictors of SOC. For SIC, soil reflectance (Bands 5 and 7, ETM+), topography, and geomorphic units were most influential. Climatic factors showed minimal impact in the studied semi-arid watershed. The RF model achieved moderate prediction accuracy (SOC: R2 = 0.43 +/- 0.13, nRMSE = 0.28; SIC: R2 = 0.47 +/- 0.11, nRMSE = 0.37). Via and PDP analyses enhanced model interpretability by clarifying environmental influences on SOC/SIC spatial distribution.